Experiment reveals how hidden agendas in AI advice can distort human decisions

Researchers at the University of Hong Kong and Tsinghua University have established, in the Proceedings of the National Academy of Sciences no less, that covertly misaligned AI advice can manipulate human preferences. Stop the presses. Peer review, that great filter of our age, has now certified the finding that advice influences the person receiving it, provided the advice arrives with a hidden agenda attached. That a laboratory called the Laboratory of Neuropsychology and Human Neuroscience was required for this is the sort of detail that makes you wonder what everyone else in the building is doing.

'Covertly misaligned' is the phrase doing the heavy lifting, and it deserves a moment of appreciation. Not lying, not deceiving, not gaming the user — misaligned, as though the poor thing merely needs its spine adjusted. Somewhere a procurement team is reading this and concluding that misalignment is a configuration issue rather than a property of systems trained on our own arguments and then pointed at people. And note who wrote the hidden agenda: a human, in a prompt, on purpose. The model is only the messenger, albeit a messenger now scaled to millions of simultaneous conversations.

All of this lands in a world where chatbots are being handed the tutoring, the therapy, the health questions and, God help us, the voting guidance. We have already written about the coming influence wave from the other direction, the one targeting the chatbots themselves, and it turns out the arrow runs both ways. The news is not that AI can shape what people think — anyone with a growth deck has known that for years. The news is that it can do so covertly, harmlessly phrased, and still walk away with a citation. Expect the mitigation to be a disclosure footer, and expect that footer to be ignored.